EUVIMEDEUVIMED entdecken
Crossref

A Deep Machine Learning-Driven Predictive Intelligence Framework for Data-Centric Early Warning and Anticipatory Detection of Emerging and Non-Traditional Security Threats in Pakistan

Zainab Akhtar Chaudhary, Muhammad Usman Malik, Zulfqar Khan

Applied Linguistics Research Review · 31.10.2026 · 5 (10(II)) · 100-131

DOI: 10.63468/alrr.114 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Applied Linguistics Research Review, online erschienen am 31.10.2026. Verfasst von Zainab Akhtar Chaudhary und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Pakistan faces a compound threat environment in which a resurgent militancy co-exists with poliovirus transmission, climate-driven disasters, macroeconomic stress, cyber-crime and large cross-border population movements. This study designs and tests a data-centric predictive intelligence framework that fuses these non-traditional security (NTS) signals for anticipatory early warning. Using a real-data annual panel for 2010–2025 compiled from 15 institutional sources (PIPS, SATP, CRSS, PICSS, IEP-GTI, WHO-GPEI, NIH, NDMA, World Bank, FIA, UNHCR-IOM), we estimate a Bayesian multi-source state-space model that reconciles definitionally inconsistent violence series into a latent Militancy Threat Factor (MTF). The factor is analysed with an econometric core sieve-bootstrap ADF, KPSS and Zivot–Andrews tests, Bai–Perron multiple breaks, a Hamilton Markov-switching model, bootstrap ARDL bounds tests, Toda–Yamamoto causality, multiple-imputation dynamic regression, quantile regression and NB2 count models and forecast with six deep architectures (deep MLP, LSTM, GRU, temporal convolutional network, Transformer encoder and Attention-LSTM) trained with sieve-bootstrap augmentation, deep ensembles and Monte-Carlo dropout under strict rolling-origin evaluation. The MTF shows three statistically dated breaks (2016, 2019, 2023; bootstrap sup-F p ≤ 0.004) and a persistent escalation regime since 2021 (expected duration ≈ 11 years). Recurrent networks achieve the lowest out-of-sample error (Attention-LSTM RMSE = 0.290 versus 0.342 for a random walk and 0.300 for AR(1)), with zero false alarms and a Brier skill score of 0.31, but Diebold–Mariano tests do not establish significant superiority on nine test years, and no model anticipated the 2021 turning point. Explainability analysis identifies inflation as the dominant learned predictor. Threats are spatially concentrated: Khyber Pakhtunkhwa and Balochistan account for 96% of 2025 attacks, and provincial militancy and polio burdens co-locate (Spearman ρ = 0.88). The paper offers a transparent, reproducible template that couples latent-state econometrics with uncertainty-aware deep learning, and it reports the evidential limits of annual data candidly positioning deep models as regime-aware decision-support tools rather than oracles.

Zitieren

APA

Chaudhary, Z. A., Malik, M. U., & Khan, Z. (2026). A Deep Machine Learning-Driven Predictive Intelligence Framework for Data-Centric Early Warning and Anticipatory Detection of Emerging and Non-Traditional Security Threats in Pakistan. Applied Linguistics Research Review, 5(10(II)), 100-131. https://doi.org/10.63468/alrr.114

Vancouver

Chaudhary ZA, Malik MU, Khan Z. A Deep Machine Learning-Driven Predictive Intelligence Framework for Data-Centric Early Warning and Anticipatory Detection of Emerging and Non-Traditional Security Threats in Pakistan. Applied Linguistics Research Review. 2026;5(10(II)):100-131. doi:10.63468/alrr.114

BibTeX
@article{Chaudhary2026,
  author = {Chaudhary, Zainab Akhtar and Malik, Muhammad Usman and Khan, Zulfqar},
  title = {A Deep Machine Learning-Driven Predictive Intelligence Framework for Data-Centric Early Warning and Anticipatory Detection of Emerging and Non-Traditional Security Threats in Pakistan},
  journal = {Applied Linguistics Research Review},
  year = {2026},
  volume = {5},
  number = {10(II)},
  pages = {100--131},
  doi = {10.63468/alrr.114}
}
Kostenlos starten →